CN113109780B - High-resolution range profile target identification method based on complex number dense connection neural network - Google Patents

Abstract

The invention discloses a high-resolution range profile target identification method based on a plurality of densely connected neural networks, which comprises the following steps: acquiring a radar range profile data set; carrying out short-time Fourier transform on the radar range profile data set to obtain a complex time-frequency spectrum data set; dividing the complex time-frequency spectrum data set into a complex time-frequency spectrum training data set and a complex time-frequency spectrum verification data set; constructing a plurality of densely connected neural networks; training a plurality of densely-connected neural networks by using a plurality of time spectrum training data sets, and verifying the trained plurality of densely-connected neural networks by using a plurality of time spectrum verification data sets to obtain a trained plurality of densely-connected neural networks; and identifying the radar range profile test data set by using the trained complex dense connection neural network to obtain a target identification result. The invention utilizes the built complex dense connection neural network to train and recognize the complex high-resolution range profile and fully utilizes the characteristic structure in the signal, thereby improving the accuracy of the recognition network.

Description

A high-resolution range image target recognition method based on complex densely connected neural network

technical field

The invention belongs to the technical field of radar, and in particular relates to a high-resolution range image target recognition method based on a complex densely connected neural network.

Background technique

Radar target recognition is to use the radar echo signal of the target to realize the determination of the target type. Broadband radar usually works in the optical region, where the target can be seen as a large number of scattered points with different intensities. The High Resolution Range Profile (HRRP) is the vector sum of the echoes of each scattering point on the target obtained by the broadband radar signal. It reflects the distribution of scattering points on the target along the radar line of sight, and contains important structural features of the target, and is widely used in the field of radar target recognition.

Extracting recognition features from high-resolution range images is an important link in radar target recognition systems. The original high-resolution range image data is a complex number, in which both amplitude and phase contain rich target detail information. Most of the traditional recognition methods directly perform the modulo operation on the original complex high-resolution range image to obtain the real range image data for recognition. In recent years, extending the traditional real number network to the complex number domain is an emerging research direction. Before the invention of this patent, in "Research on the Target Method of Radar High Resolution Range Profile Based on Complex Number Network" (Li Wanping), the complex convolutional neural network and the complex residual network applied to the high resolution range profile have been proposed, which are similar to the real number network. The recognition rate has achieved a good improvement.

However, due to the limitations of the basic network structure of convolutional neural networks and residual neural networks, when the recognition features are more complex and the number of network layers is increased, the problem of network degradation is prone to occur, which limits the performance of target recognition.

SUMMARY OF THE INVENTION

In order to solve the above problems existing in the prior art, the present invention provides a high-resolution range image target recognition method based on a complex densely connected neural network.

An embodiment of the present invention provides a high-resolution distance image target recognition method based on a complex densely connected neural network, including:

Obtain the radar range image dataset;

performing short-time Fourier transform on the radar range image data set to obtain a complex time spectrum data set;

Dividing the complex time spectrum data set into a complex time spectrum training data set and a complex time spectrum verification data set;

Build complex densely connected neural networks;

Use the complex time spectrum training data set to train a complex dense connection neural network, and use the complex time spectrum verification data set to verify the trained complex dense connection neural network to obtain a trained complex dense connection neural network;

The target recognition result is obtained by identifying the radar range image test data set by using the trained complex densely connected neural network.

In an embodiment of the present invention, dividing the complex time spectrum data set into a complex time spectrum training data set and a complex time spectrum verification data set includes:

Perform real part and imaginary part separation on each complex time spectrum data in the complex time spectrum data set to obtain a real part time spectrum data set and an imaginary time spectrum data set;

dividing the real-time spectrum data set into a real-time spectrum training data set and a real-time spectrum verification data set;

Correspondingly, the imaginary part time spectrum data set is divided into an imaginary part time spectrum training data set and an imaginary part time spectrum verification data set;

The complex time spectrum training data set is composed of the real time spectrum training data set and the imaginary time spectrum training data set, and the real time spectrum verification data set and the imaginary time spectrum verification data Sets correspond to composing the complex time-spectrum validation datasets.

In an embodiment of the present invention, the constructed complex densely connected neural network includes an input layer, a first dense block, a first vector splicing layer, a second vector splicing layer, a data output preprocessing block, and an output layer that are connected in sequence. , the first dense block is further connected to the first vector splicing layer through a second dense block, and the first vector splicing layer is further connected to the second vector splicing layer through a third dense block.

In an embodiment of the present invention, the first dense block includes a complex batch normalization layer, a complex activation function layer, a complex convolution layer, a complex batch normalization layer, a complex activation function layer, a complex convolution layer connected in sequence Laminate.

In one embodiment of the present invention, the second dense block includes a complex batch normalization layer, a complex activation function layer, a complex convolution layer, a complex batch normalization layer, a complex activation function layer, a complex convolution layer connected in sequence Laminate.

In an embodiment of the present invention, the third dense block includes a complex batch normalization layer, a complex activation function layer, a complex convolution layer, a complex batch normalization layer, a complex activation function layer, a complex convolution layer connected in sequence Laminate.

In one embodiment of the present invention, the data output preprocessing block includes a complex batch normalization layer, a complex activation function layer, a complex convolution layer, a flat one-dimensionalization layer, a fully connected layer, and a complex batch normalization layer connected in sequence. Unification layer, fully connected layer, activation function layer.

In one embodiment of the present invention, training a complex densely connected neural network using the complex time-spectral training data set includes:

The complex densely connected neural network is trained by using the real part time spectrum training data set and the imaginary part time spectrum training data set.

In one embodiment of the present invention, verifying the trained complex densely connected neural network using the complex time-spectrum verification dataset includes:

The trained complex densely connected neural network is verified using the real time-spectrum verification data set and the imaginary time-spectrum verification data set.

In an embodiment of the present invention, using the trained complex densely connected neural network to identify the radar range image test data set to obtain the target identification result includes:

Performing short-time Fourier transform on the radar range image test data set to obtain a complex time spectrum test data set;

Perform real part and imaginary part separation to obtain real part time spectrum test data set and imaginary part time spectrum test data set in each complex time spectrum test data set in the described complex number time spectrum test data set;

The target recognition result is obtained by inputting the real-part time-spectrum test data set and the imaginary-part time-spectrum test data set into the trained complex densely connected neural network for identification.

Compared with the prior art, the beneficial effects of the present invention:

The high-resolution range image target recognition method based on the complex densely connected neural network provided by the present invention can use the constructed complex densely connected neural network to train and recognize the complex high-resolution range image, and make full use of the characteristic structure in the signal, thereby improving the network performance. recognition accuracy.

The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

Description of drawings

1 is a schematic flowchart of a high-resolution range image target recognition method based on a complex densely connected neural network provided by an embodiment of the present invention;

2 is a schematic structural diagram of a complex densely connected neural network in a high-resolution range image target recognition method based on a complex densely connected neural network provided by an embodiment of the present invention;

3 is a schematic structural diagram of another complex densely connected neural network in a high-resolution range image target recognition method based on a complex densely connected neural network provided by an embodiment of the present invention;

4 is a schematic diagram of a comparison result of correct recognition rates of several target recognition methods provided by an embodiment of the present invention;

5 is a schematic diagram of the change curve of the recognition rate in all target recognition processes provided by an embodiment of the present invention;

FIG. 6 is a schematic diagram of change curves of loss values in the process of identifying all targets provided by an embodiment of the present invention.

Detailed ways

The present invention will be described in further detail below with reference to specific embodiments, but the embodiments of the present invention are not limited thereto.

Example 1

Please refer to FIG. 1. FIG. 1 is a schematic flowchart of a high-resolution range image target recognition method based on a complex densely connected neural network provided by an embodiment of the present invention. This embodiment proposes a high-resolution range image target recognition method based on a complex densely connected neural network, and the high-resolution range image target recognition method based on a complex densely connected neural network includes the following steps:

Step 1. Obtain the radar range image dataset.

Specifically, the radar high-resolution range profile (HRRP) is the vector sum of the echoes of each scattering point on the target obtained by the broadband radar signal. It reflects the distribution of the scattering points on the target along the radar line of sight and includes important structures of the target. feature. The original high-resolution range image data is complex, which not only can reflect a lot of target structure information, but also has the advantages of fast processing and easy acquisition. In this embodiment, a radar range image data set is obtained by acquiring several radar range image data.

Step 2: Perform short-time Fourier transform on the radar range image data set to obtain a complex time spectrum data set.

Specifically, in this embodiment, the radar high-resolution range image is subjected to short-time Fourier transform to obtain its complex time spectrum data, and the short-time Fourier transforms the time domain signal to the frequency domain to obtain more signal details. Before performing the short-time Fourier transform on the radar range image data set, each radar range image data in the radar range image data set can also be absolutely aligned to align the zero phase. The main operation is to convert the original data. to move the frequency point to the center of the spectrum, overcoming translation sensitivity. After the alignment processing, in order to overcome the intensity sensitivity, the intensity normalization method is adopted, and the amplitude normalization of the aligned radar range image data on the dimension k is expressed as:

表示雷达距离像数据在维度k上的归一化幅度。本实施例对上述绝对对齐、归一化得到的数据进行短时傅里叶变换得到其复数时频谱数据h，其中，h是m×n的复数矩阵，对每个绝对对齐、归一化的数据进行短时傅里叶变换得到复数时频谱数据集。短时傅立叶变换是将信号加滑动时间窗，并对窗内信号做傅立叶变换得到信号的时频谱数据，因而它的时间分辨率和频率分辨率受Heisenberg测不准原理约束，一旦窗函数选定，时频分辨率便确定下来，具体通过短时傅里叶变换得到的每个复数时频谱数据表示为：where f _k represents the magnitude of the radar range image data on dimension k,

Represents the normalized magnitude of the radar range image data in dimension k. In this embodiment, short-time Fourier transform is performed on the data obtained by the above absolute alignment and normalization to obtain the complex time spectrum data h, where h is an m×n complex number matrix. The data is subjected to a short-time Fourier transform to obtain a complex-time spectral dataset. The short-time Fourier transform is to add a sliding time window to the signal, and perform Fourier transform on the signal in the window to obtain the time spectrum data of the signal. Therefore, its time resolution and frequency resolution are constrained by the Heisenberg uncertainty principle. Once the window function is selected , the time-frequency resolution is determined. Specifically, each complex time-frequency spectrum data obtained by short-time Fourier transform is expressed as:

Among them, τ represents time, ω represents frequency, f( ) represents the absolute alignment and normalized radar range image data that needs to be subjected to short-time Fourier transform, and L( ) represents the data in the short-time Fourier transform. The Hamming window function used. For example, set the window function as a Hamming window, the overlap interval as 0, the window function length as 32, and the number of Fourier points as 16, and finally obtain complex time spectrum data with a size of 32×15.

Step 3: Divide the complex time spectrum data set into a complex time spectrum training data set and a complex time spectrum verification data set.

Specifically, each complex time spectrum data in the complex time spectrum data set obtained in step 2 of this embodiment can be expressed as h=x+yi, where x is an m×n real number matrix, and y is an m×n real number matrix, the complex time spectrum data set is divided into the complex time spectrum training data set and the complex time spectrum verification data set, including:

Separate the real and imaginary parts of each complex time spectrum data set in the complex time spectrum data set to obtain the real time spectrum data set and the imaginary time spectrum data set; divide the real time spectrum data set into real time spectrum training data sets , real part time spectrum verification data set; correspondingly, the imaginary part time spectrum data set is divided into imaginary part time spectrum training data set and imaginary part time spectrum verification data set; The data set corresponds to the complex time spectrum training data set, and the real part time spectrum verification data set and the imaginary part time spectrum verification data set correspond to form the complex time spectrum verification data set. For example, in this embodiment, three types of aircraft data are processed and divided into complex time spectrum training data set and complex time spectrum verification data set according to the ratio of 30:1, that is, the ratio of real part time spectrum training data set and real part time spectrum verification data set is: 30:1, the ratio of the corresponding imaginary part time spectrum training data set and imaginary part time spectrum verification data set is 30:1, and they are divided into batches of fixed size, each batch contains 100 signals, which are used for subsequent Training and validation of complex densely connected neural networks.

Step 4. Build a complex densely connected neural network.

公式计算得到，Var(·)表示求方差，E(·)表示求期望，可以看出，w的方差只取决于模值与相位无关，所以在复数参数初始化中，本实施例使用合适的参数为σ的瑞利分布初始化复数参数w的模值，然后再使用在(-π,π)上的均匀分布初始化相位，为了保持输入输出的方差和网络训练中的梯度一致，初始化时设

其中，N_in、N_out分别表示网络中输入、输出的样本个数。Specifically, please refer to FIG. 2. FIG. 2 is a schematic structural diagram of a complex densely connected neural network in a high-resolution range image target recognition method based on a complex densely connected neural network provided by an embodiment of the present invention. The densely connected neural network includes an input layer, a first dense block, a first vector splicing layer, a second vector splicing layer, a data output preprocessing block, and an output layer that are sequentially connected. The first dense block is also connected to the first dense block through the second dense block. The vector splicing layer is connected, and the first vector splicing layer is also connected with the second vector splicing layer through the third dense block. In this embodiment, a complex densely connected neural network with dual channels of real and imaginary parts is constructed, and the input of the complex densely connected neural network is a real part matrix x and an imaginary part matrix y. The two-channel complex densely connected neural network is built, and each dense block is directly connected by a line, so that there is a short path directly connected from the early layer to the later layer, which can maximize the flow of information between network layers and prevent The degradation problem of complex densely connected neural networks. The complex densely connected neural network receives the real part matrix x and imaginary part matrix y of the signal h=x+yi, and the initial complex convolution kernel w=a+bi in the network is also divided into real part matrix a and imaginary part matrix b, respectively and The real part matrix and the imaginary part matrix of the signal received by the network are convolved, and the final convolution result is w*h=(a*xb*y)+i(b*x+a*y). Among them, the complex parameters initialized in the complex densely connected neural network can be expressed as the product of the modulus and the phase w=|w|e ^iθ , and the variance of w can be obtained by

Calculated by the formula, Var( ) represents the variance, and E( ) represents the expectation. It can be seen that the variance of w only depends on the modulus value and has nothing to do with the phase. Therefore, in the initialization of complex parameters, this embodiment uses appropriate parameters Initialize the modulo value of the complex parameter w for the Rayleigh distribution of σ, and then use the uniform distribution on (-π, π) to initialize the phase. In order to keep the variance of the input and output consistent with the gradient in the network training, set

Among them, N _in and N _out represent the number of input and output samples in the network, respectively.

Further, please refer to FIG. 3. FIG. 3 is a schematic structural diagram of another complex densely connected neural network in a high-resolution range image target recognition method based on a complex densely connected neural network provided by an embodiment of the present invention. A dense block includes a complex batch normalization layer, a complex activation function layer, a complex convolution layer, a complex batch normalization layer, a complex activation function layer, and a complex convolution layer connected in sequence, such as the activation function of the complex activation function layer. All are CReLU activation functions. The complex convolution layers have 32 convolution kernels. The size of each convolution kernel is 1*3 and the convolution stride is 2. The complex batch normalization layer is normalized to BN. Normalization function. In this embodiment, the first dense block performs the above-mentioned sequential connection layer processing on the real-part time-spectrum training data set and the imaginary-part time-spectrum training data set input in step 3 . Among them, for the complex batch normalization layer, in order to normalize each complex time spectrum training data to obey the standard normal distribution, this embodiment first normalizes the complex time spectrum training data h as:

Cov(·)表示求取协方差，E(·)表示求期望，R(·)表示取复数数据的实部，S(·)表示取复数数据的虚部。在归一化后，与实数数据的批归一化类似，我们也引入尺度系数和偏置两个变量，最终的复数批归一化公式为

其中，

表示归一化后的复数时频谱训练数据；γ代表尺度系数，为一大小为2×2的半正定矩阵，β表示偏置参数，为一具有两个学习度(虚部和实部)的复数参数。在初始化中，因为复数数据

的实部和虚部方差都是1，所以设

γ_ri＝0，β＝0。where V represents the 2×2 covariance matrix,

Cov(·) represents finding covariance, E(·) represents finding expectation, R(·) represents finding real part of complex data, and S(·) represents finding imaginary part of complex data. After normalization, similar to batch normalization of real data, we also introduce two variables, scale coefficient and bias, and the final complex batch normalization formula is

in,

represents the normalized complex time spectral training data; γ represents the scale coefficient, which is a positive semi-definite matrix of size 2 × 2, and β represents the bias parameter, which is a matrix with two learning degrees (imaginary part and real part). Plural parameters. in initialization, because complex data

The variances of the real and imaginary parts are both 1, so let

γ _ri =0, β=0.

For the complex activation function layer, the ReLU function is used in the real and imaginary parts respectively. When the real and imaginary parts of the complex data are greater than 0 or less than 0 at the same time, the activation function satisfies the Cauchy-Riemann equation. The specific activation function is expressed as :

CReLU(z)=ReLU(R(z))+iReLU(S(z));

Among them, R(·) represents taking the real part of complex data, and S(·) represents taking the imaginary part of complex data.

For the complex convolution layer, the real part and the imaginary part are regarded as two real parts with different logical meanings, and the two-dimensional convolution layer is input through different channels respectively, and the complex data is abstractly expressed as:

h=x+yi;

Among them, h represents the whole complex data, in this embodiment is the input complex time spectral data of the complex densely connected neural network for each complex time spectral data set, x represents the real part of the complex data, y represents the imaginary part of the complex data, i represents the imaginary unit. In order to make the convolution operation in the complex domain to achieve the same effect as the real domain, the used convolution kernel is also a complex filter matrix expressed as:

w=a+bi;

Among them, a represents the real part matrix, b represents the imaginary part matrix, and i represents the imaginary part unit. The specific operation of complex convolution is that the four parts of the real and imaginary parts of the complex number and the real and imaginary parts of the convolution kernel are cross-convolved in pairs. The real part takes the result of subtraction, and the imaginary part takes the result of addition. The calculation formula is expressed as:

w*h=(a*x-b*y)+(b*x+a*y)i;

Rewrite the above formula in matrix form as:

Among them, R(·) represents taking the real part of complex data, and S(·) represents taking the imaginary part of complex data.

Further, referring to FIG. 3 again, the second dense block in this embodiment includes a complex batch normalization layer, a complex activation function layer, a complex convolution layer, a complex batch normalization layer, a complex activation function layer, a complex number activation function layer, and a complex number batch normalization layer. convolutional layer. In this embodiment, the second dense block performs the above-mentioned sequential connection layer processing on the complex data set output by the first dense block processing. The implementation of the specific connection layer is similar to that of the first dense block processing, which is not repeated here.

Further, referring to FIG. 3 again, the third dense block of the present embodiment includes a complex batch normalization layer, a complex activation function layer, a complex convolution layer, a complex batch normalization layer, a complex activation function layer, a complex number activation function layer, and a complex number batch normalization layer. convolutional layer. In the third dense block in this embodiment, after the complex data set output by processing the second dense block and the complex data set output by processing the first dense block are spliced through the first vector splicing layer, the above-mentioned sequential connection layer processing is performed. , the implementation of the specific connection layer is similar to the processing of the first dense block and the second dense block, which is not repeated here.

Further, please refer to FIG. 3 again, the data output preprocessing block of this embodiment includes a complex batch normalization layer, a complex activation function layer, a complex convolution layer, a flat one-dimensional layer, a fully connected layer, and a complex batch that are sequentially connected. Normalization layer, fully connected layer, activation function layer. In the data output preprocessing block in this embodiment, after the complex data set output by the third dense block processing and the complex data set output by the second dense block processing are spliced through the second vector splicing layer, the above sequential connection layer is performed. Processing, the implementation of the complex batch normalization layer, complex activation function layer, and complex convolution layer in the specific connection layer is similar to the processing of the first dense block, the second dense block, and the third dense block, as well as the flat one-dimensional layer, the full The connection layer is conventional processing, the fully connected layer connected with the flat one-dimensional layer reduces the dimension to 300, the fully connected layer connected with the complex batch normalization layer reduces the dimension to 3, and the data output in the preprocessing block is used as the output The activation function of the activation function layer is the softmax activation function, which is the type of the vector with the dimension of 3, and the output of the classification. The specific implementation of each layer will not be repeated here.

Step 5, using the complex time spectrum training data set to train the complex densely connected neural network, and using the complex time spectrum verification data set to verify the trained complex densely connected neural network to obtain a trained complex densely connected neural network.

Specifically, the present embodiment uses the complex time spectrum training data set to train the complex densely connected neural network. Since the complex time spectrum training data set includes the real time spectrum training data set and the imaginary time spectrum training data set, the real time spectrum training data set is used. The spectral training dataset, the imaginary time spectral training dataset trains a complex densely connected neural network. This embodiment uses the real part time spectrum training data set and the imaginary part time spectrum training data set to accurately train the complex densely connected neural network, so that the complex densely connected neural network is gradually fitted, and different structural features in the signal can be accurately extracted.

Further, the present embodiment utilizes the complex time-spectrum verification data set to verify the training of the complex densely connected neural network, also due to the real-time spectrum verification data set and the imaginary-part time-spectrum verification data set, the real-time spectrum verification data set, The imaginary time-spectral validation dataset validates the trained complex densely connected neural network. Each training of all real-time spectrum training datasets and imaginary-time spectrum training datasets, correspondingly use real-time spectrum validation datasets and imaginary-time spectrum validation datasets for validation. If the validation result is a drop, then It shows that the network training has over-fitting, indicating that the network training parameters need to be adjusted. After many adjustments, the optimal network parameters of the complex densely connected neural network in this embodiment are finally determined, and finally a trained complex densely connected neural network is obtained.

In this embodiment, in order to make more use of the information in the original complex radar range image data, because the complex radar range image data is directly used, the amount of data calculation in the network will be very large, which will affect the learning speed of the network. The data information of the complex radar range image and the learning speed of the network are used as much as possible, so the real part and the imaginary part of the radar range image data are separately input into the dual-channel network, that is, the complex densely connected neural network is separately input for training.

Step 6: Use the trained complex densely connected neural network to identify the radar range image test data set to obtain the target identification result.

Specifically, in this embodiment, the trained complex densely connected neural network is used to identify the radar range profile test data set to obtain the target identification result. For any radar range profile test data, similar processing is required for the radar range profile training data set and the radar range profile validation data set, specifically: performing short-time Fourier transform on the radar range profile test data set to obtain a complex time spectrum test Data set; separate the real part and imaginary part of each complex time spectrum test data set in the complex time spectrum test data set to obtain the real part time spectrum test data set and the imaginary part time spectrum test data set; The imaginary part time spectrum test data set is input to the trained complex densely connected neural network for recognition to obtain the target recognition result. Among them, the specific process of performing short-time Fourier transform on the complex time spectrum test data set, and the specific process of separating the real part and imaginary part of the complex time spectrum test data in the complex time spectrum test data set, and the radar distance image training data set, radar The distance is similar to the processing of the validation dataset and will not be repeated here.

In order to verify the effectiveness of the high-resolution range image target recognition method based on the complex densely connected neural network proposed in this embodiment, the following experiments on measured data are further explained.

1. Experiment content

In this experiment, high-resolution range images of three types of aircraft targets are used to train the recognition system. The parameters of the three types of aircraft targets and the radar parameters of the high-resolution range images of the three types of aircraft targets are shown in Table 1.

Table 1 Parameters and radar parameters of three types of aircraft targets

In Table 1, the "Yacques-42" aircraft target contains seven segments of high-resolution range image data, the "An-26" aircraft target contains seven segments of high-resolution range image data, and the "Citation" aircraft target contains five segments of high-resolution range image data. In this experiment, the second and fifth high-resolution range image data of the "Yacques-42" aircraft target, the sixth and seventh high-resolution range image data of the "Citation" aircraft target, and the "An-26" aircraft target were selected. The fifth and sixth sections of high-resolution range image data are used as training samples for training the recognition system, and the remaining data are used as test samples for testing the performance of the recognition system. All high-resolution range image data are 256 range units.

At the same time, under the test data of the above-mentioned three types of aircraft targets in this embodiment, the complex densely connected neural network based on the time spectrum of the present invention and the traditional high-resolution range image-based classical real convolutional neural network and classical real densely connected neural network are respectively used. , A total of four methods of complex convolutional neural network are used for experimental test comparison.

2. Analysis of experimental results

Please refer to FIG. 4. FIG. 4 is a schematic diagram of the comparison results of the correct recognition rates of several target recognition methods provided by the embodiment of the present invention. It can be seen from FIG. 4 that the recognition rate of the time-spectrum-based complex densely connected neural network is about 98.21 %, significantly higher than 94.6% of the classical real number convolutional neural network based on high-resolution range profile and 94.19% of the classical real number densely connected neural network based on high-resolution range profile, and 94.19% of the high-resolution range profile-based complex convolutional neural network. Compared with the recognition rate of 97.6%, it has also been significantly improved.

Please refer to FIG. 5 and FIG. 6 . FIG. 5 is a schematic diagram of the change curve of the recognition rate in the process of recognizing all targets provided by the embodiment of the present invention, and FIG. 6 is the change curve of the loss value in the process of recognizing all the targets provided by the embodiment of the present invention. It can be seen from Figure 5 and Figure 6 that this embodiment no longer directly modulates the original signal to obtain a real number signal, but uses the real part and imaginary part of the original signal respectively, and in each layer of the network structure A direct connection is built, thereby retaining more target structure information in the original signal, which helps to improve the recognition accuracy of the network.

To sum up, the high-resolution range image target recognition method based on the complex densely connected neural network provided in this embodiment can use the constructed complex densely connected neural network to train and recognize the complex high-resolution range image. The neural network can better cope with the degradation problem because of the direct connection structure, and the recognition network in this embodiment is based on time spectrum. Compared with the recognition network based on high-resolution range images, it can recognize more features and details The feature structure in the signal is used to improve the accuracy of the recognition network.

The above content is a further detailed description of the present invention in combination with specific preferred embodiments, and it cannot be considered that the specific implementation of the present invention is limited to these descriptions. For those of ordinary skill in the technical field of the present invention, without departing from the concept of the present invention, some simple deductions or substitutions can be made, which should be regarded as belonging to the protection scope of the present invention.

Claims (9)

1. a high-resolution distance image target recognition method based on complex number densely connected neural network, is characterized in that, comprises: Obtain the radar range image dataset; performing short-time Fourier transform on the radar range image data set to obtain a complex time spectrum data set; Dividing the complex time spectrum data set into a complex time spectrum training data set and a complex time spectrum verification data set; Build complex densely connected neural networks; Use the complex time spectrum training data set to train a complex dense connection neural network, and use the complex time spectrum verification data set to verify the trained complex dense connection neural network to obtain a trained complex dense connection neural network; Using the trained complex densely connected neural network to identify the radar range image test data set to obtain the target identification result; The constructed complex densely connected neural network includes an input layer, a first dense block, a first vector splicing layer, a second vector splicing layer, a data output preprocessing block, and an output layer that are connected in sequence. The first dense block It is also connected with the first vector stitching layer through a second dense block, and the first vector stitching layer is also connected with the second vector stitching layer through a third dense block. 2. the high-resolution distance image target recognition method based on the complex number densely connected neural network according to claim 1, is characterized in that, the spectrum data set when the complex number is divided into the spectrum training data set when the complex number, the spectrum verification data when the complex number is divided Sets include: Perform real part and imaginary part separation on each complex time spectrum data in the complex time spectrum data set to obtain a real part time spectrum data set and an imaginary time spectrum data set; dividing the real-time spectrum data set into a real-time spectrum training data set and a real-time spectrum verification data set; Correspondingly, the imaginary part time spectrum data set is divided into an imaginary part time spectrum training data set and an imaginary part time spectrum verification data set; The complex time spectrum training data set is composed of the real time spectrum training data set and the imaginary time spectrum training data set, and the real time spectrum verification data set and the imaginary time spectrum verification data Sets correspond to composing the complex time-spectrum validation datasets. 3. The high-resolution distance image target recognition method based on a complex densely connected neural network according to claim 1, wherein the first dense block comprises a complex batch normalization layer, a complex activation function layer, Complex convolutional layer, complex batch normalization layer, complex activation function layer, complex convolutional layer. 4. The high-resolution distance image target recognition method based on a complex densely connected neural network according to claim 3, wherein the second dense block comprises a complex batch normalization layer, a complex activation function layer, a complex activation function layer that are sequentially connected. Complex convolutional layer, complex batch normalization layer, complex activation function layer, complex convolutional layer. 5. The high-resolution distance image target recognition method based on a complex densely connected neural network according to claim 3, wherein the third dense block comprises a complex batch normalization layer, a complex activation function layer, Complex convolutional layer, complex batch normalization layer, complex activation function layer, complex convolutional layer. 6. The high-resolution distance image target recognition method based on a complex densely connected neural network according to claim 3, wherein the data output preprocessing block comprises a complex batch normalization layer, a complex activation function layer connected in turn , complex convolution layer, flat one-dimensional layer, fully connected layer, complex batch normalization layer, fully connected layer, activation function layer. 7. the high-resolution distance image target recognition method based on the complex number densely connected neural network according to claim 2, is characterized in that, utilizing the described complex number time spectrum training data set to train the complex number densely connected neural network comprises: The complex densely connected neural network is trained by using the real part time spectrum training data set and the imaginary part time spectrum training data set. 8. the high-resolution distance image target recognition method based on the complex number densely connected neural network according to claim 2, is characterized in that, the complex number densely connected neural network that utilizes the described complex number time spectrum verification data set to verify the training comprises: The trained complex densely connected neural network is verified using the real time-spectrum verification data set and the imaginary time-spectrum verification data set. 9. the high-resolution range image target recognition method based on the complex number dense connection neural network according to claim 1, is characterized in that, utilizes described trained complex number dense connection neural network to identify the radar range image test data set to obtain the target The identification results include: Performing short-time Fourier transform on the radar range image test data set to obtain a complex time spectrum test data set; Perform real part and imaginary part separation to obtain real part time spectrum test data set and imaginary part time spectrum test data set in each complex time spectrum test data set in the described complex number time spectrum test data set; The target recognition result is obtained by inputting the real-part time-spectrum test data set and the imaginary-part time-spectrum test data set into the trained complex densely connected neural network for identification.

Images